Cloud computing resource load prediction based on improved informer

Load prediction is a crucial task in the area of cloud computing resource management. Due to the ever-changing characteristics of the workload data in Kubernetes clusters, influenced by time and varying access patterns of different applications, these uncertainties present significant challenges to the time series prediction of workload data. With the rapid advancement of artificial intelligence technology, deep learning methods based on the Transformer architecture have begun to gradually replace traditional statistical approaches and machine learning algorithms for time series forecasting tasks. Although the Transformer-based Informer model has achieved commendable results in time series forecasting, it fails to account for the potential distributional discrepancies between model training data and real-world testing data in load prediction scenarios, leading to insufficient model generalization. This limitation results in suboptimal prediction accuracy when applied to load forecasting in production environments. To enhance the precision of load forecasting, this paper proposes a load prediction model based on an improved Informer model. By introducing an uncertainty domain shift modeling module during the training process of the Informer model, we establish a Gaussian distribution for the feature statistics of the training data and randomly sample new feature statistics from this distribution. This process leverages randomness to simulate the potential domain shifts and uncertainties in data distribution on the testing set, thereby improving the model's generalization capability and forecast accuracy. Experimental results indicate that the improved Informer model proposed in this paper achieves more desirable predictive performance in the context of cloud computing resource load prediction.